feat: accumulate function_call events in streaming path

[ashwing]

Summary

The streaming path's ResponseAccumulator handles text message events but didn't yet have handlers for function_call events — they fell through the wildcard arm. This adds the missing accumulation so streaming responses include FunctionCall output items, matching the blocking JSON path.

Prerequisite for #51 (execute_loop needs FunctionCall items available in the streaming path to dispatch tools).

Before

A client sends a streaming request with tools:

{"model": "meta-llama/...", "stream": true, "input": "What's the weather?", "tools": [{"type": "function", "name": "get_weather", ...}]}

vLLM responds with function_call SSE events, but the accumulator ignores them. The client gets back:

{"status": "completed", "output": []}

After

The same request now produces:

{
  "status": "completed",
  "output": [{
    "type": "function_call",
    "id": "fc_1",
    "call_id": "call_abc",
    "name": "get_weather",
    "arguments": "{\"location\":\"Paris\"}",
    "status": "completed"
  }]
}

What's handled

• OutputItemAdded with item_type == "function_call" — starts a new in-flight FunctionToolCall
• FunctionCallArgumentsDelta — appends to argument buffer
• FunctionCallArgumentsDone — finalizes with authoritative arguments, pushes to output
• Multiple function calls in one response (parallel tool use)
• Interleaved message → function_call → message ordering preserved
• Stream disconnect mid-arguments — partial args retained via forced finalize
• Orphaned deltas (no active function call) — safely cleared before next call
• Coexists with reasoning accumulation (PR feat: support reasoning output items #57) — unified match with three-way type dispatch

Test Plan

Unit tests (8 in accumulator.rs):

• Full lifecycle, parallel tool use, interleaved with messages, orphaned deltas, forced finalize on disconnect

Cassette integration tests (12 in accumulator_cassette_test.rs):

• tool_choice=auto streaming + non-streaming (parallel tool calls)
• tool_choice=required streaming + non-streaming
• tool_choice=named streaming + non-streaming
• tool_choice=none streaming + non-streaming (zero function calls)
• Reasoning streaming — Qwen3 (reasoning + message)
• Reasoning streaming — GPT-oss (reasoning only, see note below)
• Legacy Gemma4 function_call + text-only regression guard

Non-streaming tests exercise the from_json path against the same cassettes.

All tests pass, cargo fmt --check + cargo clippy -- -D warnings clean.

Note: GPT-oss streaming gap

While testing the reasoning cassettes, I found that GPT-oss doesn't emit response.output_item.added for the message item after reasoning — it jumps from reasoning_text.done → output_text.done → output_item.done. The streaming accumulator can't capture that message because no output_item.added creates the in-flight state. Qwen3 emits the event correctly. Not blocking — the full output is in the response.completed payload. Filed as #62 for follow-up.

[ashwing]

@maralbahari This is the function_call accumulation PR we discussed in #51. Ready for review when you get a chance.

Covers the full streaming lifecycle: OutputItemAdded(function_call) → argument deltas → FunctionCallArgumentsDone → finalized OutputItem::FunctionCall in the response payload. Also handles parallel tool calls, interleaving with text messages, and stream disconnect with partial args.

Once this lands, #51's dispatch_tools will have real FunctionCall items to work with in the streaming path.

[maralbahari]

@ashwing thank you for the quick PR.
I have added #60 to capture all the scenarios of request payload for variety of function_call options as tool_choices to vLLM upstream and recorded them. so we can cover comprehensive tests. could you please try them out.

[ashwing]

@maralbahari thanks! I rebased on top of #60 and added cassette tests that exercise all 4 tool_choice modes through the accumulator:

• auto streaming + non-streaming (parallel tool calls — get_stock_price + search_web)
• required streaming + non-streaming
• named streaming + non-streaming
• none streaming + non-streaming (asserts zero function calls)

Also added reasoning cassette tests (Qwen3 + GPT-oss) as a regression guard — found an interesting edge case with GPT-oss where it skips output_item.added for the message after reasoning. Filed #62 for follow-up.

12 cassette tests total, all passing.